The impressive capabilities of artificial intelligence are demonstrated by models like ChatGPT and other deep generative models (DGMs). These barrier-breaking technologies can autonomously learn from a vast amount of samples and generate something new that resembles what they have already learned, from poems to complete symphonies and the creation of new videos and images.
The capabilities of these models clearly show up in standard tasks. But, as a new study from the Massachusetts Institute of Technology (MIT) points out, similarity alone is not enough when it comes to mastering binary processes or engineering design problems.
According to Lyle Regenwetter, author of the study and a doctoral student in the Department of Mechanical Engineering at MIT, “Although DTMs have a lot of potential, they are flawed in their basic structure. The goal of these models is to recreate a set of data. But we engineers and designers often want to create something completely new, not just reproduce what already exists.” Regenwetter and his colleagues argue that reimagining models is necessary to benefit from Artificial Intelligence in generating new, innovative ideas and designs.
Faez Ahmed, a co-author of the study and an assistant professor of mechanical engineering at MIT, further explains this point: “Often, the effectiveness of these models is measured by the statistical similarity of the sample they generate to the pre-existing model. But in design, it’s often more important to be different and innovative, rather than just similar.”
Ahmed and Regenwetter’s study sheds light on the weaknesses of deep generative models in engineering design problems. Using a case study in bicycle frame design, they were able to show that while the frames created were very similar to previous designs, they failed in technical performance and requirements. However, when you used DGMs with engineering-oriented goals, rather than just focusing on statistical similarity, they were able to produce better frames.
The results of the study illustrate that Artificial Intelligence models on similarity cannot be transferred when applied to technical tasks, or only in a very limited way. However, the researchers also show how AI, when properly oriented, can be used effectively to assist in design activities.
Regenwetter comments, “We are trying to find a way to use AI to help engineers create innovative products faster and better. To do this, we first need to clarify the exact requirements. Our study sees itself as a first step in that direction.”
The results of this fascinating study, which was conducted in collaboration between the MIT-IBM Watson AI Lab and MIT’s DeCoDe Lab, were recently published online and can be found in the December issue of the journal Computer Aided Design.
The study again clearly emphasizes the difficulties of using DGMs for engineering tasks, explaining that standard DGMs usually do not address specific design requirements. The team used bicycle frame design as a simple case study and showed that problems can arise even in the initial learning phase.
The emphasis on statistical similarity can lead to two frames being deemed to perform the same based only on their dimensions. However, even minor differences can result in one frame being significantly weaker relative to the other. This discrepancy is often overlooked by the AI in the process.
The researchers continued the study with models designed specifically for engineering tasks. A model that aimed to prioritize statistical similarity as well as functional performance actually produced designs that were both realistic and performed better than existing designs. However, in doing so, a large number of designs were physically “invalid” because the components did not fit together properly or contained impossible overlaps.
Regenwetter commented, “We saw designs that were significantly better than the original template, but also geometrically incompatible designs because the model wasn’t set up to maintain boundaries.”
Another model, specifically designed to generate new geometric structures and prioritize physically feasible designs, produced the best-performing designs. Ahmed noted, “A model that goes beyond statistical similarity can produce designs that are better than those that already exist. This is a testament to what AI can do when explicitly trained on a specific design activity.”
The study concludes by suggesting that generative AI models should be built with other priorities in mind, such as performance, design constraints and novelty. Ahmed sees particular potential in technical areas, such as molecular design and civil infrastructure. “By pointing out the potential pitfalls of focusing on statistical similarity, we hope to stimulate new thinking and strategies for generative AI applications outside of multimedia.”
